Observables of the universal continuum tree of states in mean-field spin glasses: moments, typical values of divergent observables, and controlled approximations

Below its critical temperature, the mean-field (Sherrington–Kirkpatrick) spin glass is described by Parisi's full replica symmetry breaking solution: the equilibrium measure splits into infinitely many pure states organized in an ultrametric tree whose leaves carry statistical weights $w_\alpha$ and whose branchings occur at every level of a continuous hierarchy. Parisi (1993) constructed the probabilistic law of this continuum tree, showed that it is universal (independent of the microscopic construction), derived its branching statistics — bifurcations and trifurcations both occur at leading order in the level spacing — and obtained low-level approximations for the weight density and for the moments of $Y(x)=\sum_c w_c^2$, while leaving the observables of the tree analytically uncontrolled: closed forms for general moments and full distributions of tree-dependent observables, the typical values of observables whose ensemble averages diverge (e.g. $\sum_{\alpha,\gamma} w_\alpha w_\gamma f(x_{\alpha\gamma})$ with $f(x)\sim x^{-\lambda}$, $1<\lambda<2$), controlled approximations uniform over the full weight domain, and quantitative convergence estimates for finite cutoffs. Later work resolved the numerical generation of the tree (Parisi, Ricci-Tersenghi, Yllanes, 2015) and established the tree's structure rigorously (Ruelle's probability cascades; the Bolthausen–Sznitman coalescent; Panchenko's proof of the Parisi ultrametricity conjecture). The open core recorded here is the quantitative analytic theory of the tree's observables: their moments and distributions at general levels, their typical values in the singular regimes where averages diverge, weight-domain-uniform approximation schemes, and convergence rates for finite-resolution constructions.

Background

In a spin glass, magnetic moments interact through random couplings of fluctuating sign, so that no periodic magnetic order is possible; the mean-field archetype is the Sherrington–Kirkpatrick (SK) model, an Ising model on a complete graph with independent random couplings. Its low-temperature phase is described by Parisi's replica symmetry breaking (RSB) theory: the thermal equilibrium (Gibbs) measure splits into infinitely many pure states — ergodic valleys separated by free-energy barriers that diverge with system size — each state $\alpha$ carrying a statistical weight $w_\alpha$ proportional to $\exp(-\beta F_\alpha)$, where $F_\alpha$ is the state's free energy. Pairs of states are characterized by their overlap $q_{\alpha\beta}$, a normalized inner product of their magnetization profiles.

A central prediction of RSB theory is ultrametricity: overlaps satisfy $q_{\alpha\beta}\ge\min(q_{\alpha\gamma},q_{\gamma\beta})$, so the states organize into a hierarchical (taxonomic) tree. The tree is labeled by a level variable $x\in[0,x_M]$, where $x(q)=\int_0^q p(q')\,dq'$ is the cumulative of the disorder-averaged overlap distribution $p(q)$ and $x_M=x(q_{EA})$, where $q_{EA}$ is the self-overlap of a pure state (the Edwards-Anderson parameter, the largest overlap in the support); two states drawn independently with the weight measure have the level $x_{\alpha\gamma}$ of their closest common ancestor distributed uniformly on $[0,x_M]$. In the full (continuous) RSB solution the tree branches at every level, and the number of states with weight greater than $p$ grows as $p^{-x_M}$, i.e. the ranked weights have power-law statistics of the kind described by the Poisson–Dirichlet distribution. The mathematical counterpart of this object is Ruelle's probability cascade, a recursive construction in which independent Poisson-point-process free energies are assigned at each level; the whole law of the tree is fixed by the single Parisi function $q(x)$ of the model.

Parisi (1993) reviewed and extended the constructive description of this tree of states. He proved the equivalence of two natural hierarchical constructions of the tree measure (one assigning independent random free energies at each level, the other assigning correlated cluster weights level by level), computed the joint distribution of subcluster weights, and showed that in the continuum limit the elementary branchings are not purely binary: at level spacing $\delta$, bifurcations and trifurcations (and higher $n$-furcations, with probability scaling as $x^{n-1}$) all occur at order $\delta$, so multi-way branchings are as relevant as binary ones away from the top of the tree. In a bifurcation-only, small-$x$ approximation he derived the splitting kernel $b(z)=1/(z(1-z))$ (the density with which a cluster splits into two subclusters of relative weights $z$ and $1-z$) and solved the resulting evolution equation for the single-weight density $p_1(w,x)$ by Mellin transform (the Mellin transform of a density $p$ on $(0,1)$ is $\int_0^1 w^s p(w)\,dw$), reproducing the correct leading singularity $(1-w)^{-x}$ as $w\to 1$. He also computed the small-$x$ moments of $Y(x)=\sum_c w_c^2$, the sum of squared weights of the clusters at level $x$, finding $\langle Y(x)\rangle=1-x$ and agreement with the earlier expansions of Mézard, Parisi and Virasoro through moment order $k=6$. Finally, he argued — by mapping the construction onto random directed polymers on Cayley trees (random-energy paths on a regular tree), where Derrida and Spohn had shown that the endpoint-weight statistics depend only on an emergent power-law exponent — that the continuum tree law is universal, i.e. independent of the microscopic substratum (coordination number, free-energy distribution) used to generate it, and he gave a prescription for numerical generation in which states with weight below a cutoff $\varepsilon$ are pruned, with lost probability of order $\varepsilon^{1-x_M}$.

Parts of this picture are classical or now rigorous: the marginal statistics of the weights at any fixed level go back to Mézard, Parisi and Virasoro (1987); Ruelle (1987) and Bolthausen and Sznitman (1998) construct the continuum tree law rigorously, top-down as a cascade and bottom-up as a coalescent; and Panchenko (2013) proved the Parisi ultrametricity conjecture — the thermodynamic-limit Gibbs measure of the SK and mixed $p$-spin models is ultrametric — which pins the physical tree of states to the cascade structure. What has remained without a quantitative theory is the set of observables of the continuum tree: functionals such as $A_f=\sum_{\alpha,\gamma} w_\alpha w_\gamma f(x_{\alpha\gamma})$ built from the weights and the branching levels. Their general moments and full distributions are not known in closed form beyond low-order and small-$x$ expansions; for singular $f$ the ensemble average over the tree diverges and the physically meaningful typical value has never been computed; and the bifurcation-only approximations are controlled only near $w\to 1$ and fail at small $w$, where multi-furcations contribute at leading order.

Problem Statement

Within the framework of Parisi (J. Stat. Phys. 72, 857, 1993) — the universal continuum tree of states of a full replica symmetry breaking mean-field spin glass, whose statistics are fixed by the model's Parisi function $q(x)$, with leaf weights $w_\alpha$, cluster levels $x\in[0,x_M]$, branchings of all orders occurring at order $\delta$ in the level spacing, and a weight cutoff $\varepsilon$ under which states with $w<\varepsilon$ are pruned — compute, with analytic control, the observables of the tree. Specifically:

(1) Moments and distributions. Derive closed-form (or exact integral/Mellin-transform) expressions for the general moments $\langle A_f^k\rangle$ and the full probability distribution of the tree-dependent observable $A_f$ for general test functions $f$

$$A_f=\sum_{\alpha,\gamma} w_\alpha w_\gamma f(x_{\alpha\gamma})$$

for all levels of the hierarchy, including the contributions of trifurcations and higher-order branchings, beyond the small-$x$, at-most-one-bifurcation regime in which the moments of $Y(x)=\sum_c w_c^2$ were computed, and beyond the classical single-level weight marginals.

(2) Typical values of observables with divergent averages. For singular functions, e.g. $f(x)\sim x^{-\lambda}$ with $1<\lambda<2$, the ensemble average of $A_f$ over the tree law diverges while a typical (finite-cutoff) tree yields a finite value. Compute the typical value of $A_f$, defined as the quenched scale $\exp\langle\ln A_f\rangle$, where $\langle\ln A_f\rangle$ is the average over tree realizations of the logarithm (which remains finite where the average of the observable itself diverges); in addition, the distribution tails of $A_f$, the typical-versus-average discrepancy (between $\exp\langle\ln A_f\rangle$ and the divergent ensemble average of $A_f$), and the dependence on the cutoff $\varepsilon$ must be characterized.

(3) Controlled full-domain approximations. Construct approximation schemes for the weight statistics — e.g. the single-cluster weight density $p_1(w,x)$ — that are uniformly controlled over the full domain $0<w<1$ at general level $x$: the bifurcation-only, low-$x$ expansion reproduces the leading singularity $(1-w)^{-x}$ near $w\to 1$ but fails at small $w$, where trifurcations and higher-order processes contribute at leading order.

(4) Quantitative convergence control. Establish convergence rates and uniform error estimates for finite-resolution approximations of the continuum tree law — discrete-level constructions with level spacing $\delta x\to 0$ and pruned constructions with weight cutoff $\varepsilon\to 0$ — over the natural classes of microscopic constructions. The existence, uniqueness, and physical reality of the continuum law are already established; the quantitative finite-resolution bounds are not.

The common objective uniting these deliverables is to turn the 1993 construction of the universal continuum tree law into a quantitative theory of its observables: what a functional of the tree equals on average, what it equals typically, how its distribution behaves in the singular regimes, and how fast finite approximations converge to these values. The numerical generation of the tree itself, which the 1993 paper also left open, was solved subsequently by Parisi, Ricci-Tersenghi and Yllanes (2015) and is here a validation tool, not a target.

All four parts belong to this single objective: each is independently meaningful, and the objective is a quantitative theory of the tree's observables covering all four parts.

References

1. G. Parisi, 'On the branching structure of the tree of states in spin glasses', Journal of Statistical Physics 72 (5-6), 857-878 (1993). DOI: 10.1007/BF01048182. https://doi.org/10.1007/bf01048182
2. M. Mézard, G. Parisi, M. A. Virasoro, 'Spin Glass Theory and Beyond', World Scientific, Singapore, 1987. DOI: 10.1142/0271, ISBN 978-9971-50-116-7. https://doi.org/10.1142/0271
3. B. Derrida, H. Spohn, 'Polymers on disordered trees, spin glasses, and traveling waves', Journal of Statistical Physics 51, 817-840 (1988). DOI: 10.1007/BF01014886. https://doi.org/10.1007/BF01014886
4. D. Ruelle, 'A mathematical reformulation of Derrida's REM and GREM', Communications in Mathematical Physics 108, 225-239 (1987). DOI: 10.1007/BF01210613. https://doi.org/10.1007/BF01210613
5. E. Bolthausen, A.-S. Sznitman, 'On Ruelle's probability cascades and an abstract cavity method', Communications in Mathematical Physics 197, 247-276 (1998). DOI: 10.1007/s002200050450. https://doi.org/10.1007/s002200050450
6. D. Panchenko, 'The Parisi ultrametricity conjecture', Annals of Mathematics 177 (1), 383-393 (2013). DOI: 10.4007/annals.2013.177.1.8, arXiv:1112.1003. https://arxiv.org/abs/1112.1003
7. G. Parisi, 'Course 3: Mean field theory of spin glasses: statics and dynamics', in 'Complex Systems', Les Houches lecture notes, Elsevier, pp. 131-178 (2007); preprint version titled 'Mean field theory of spin glasses: statics and dynamics', arXiv:0706.0094. DOI: 10.1016/S0924-8099(07)80010-2. https://arxiv.org/abs/0706.0094
8. G. Parisi, F. Ricci-Tersenghi, D. Yllanes, 'Explicit generation of the branching tree of states in spin glasses', Journal of Statistical Mechanics: Theory and Experiment 2015, P05002 (2015). DOI: 10.1088/1742-5468/2015/05/P05002, arXiv:1409.8438. https://arxiv.org/abs/1409.8438
9. G. Parisi, 'The marginally stable Bethe lattice spin glass revisited', Journal of Statistical Physics 167, 515-542 (2017). DOI: 10.1007/s10955-017-1724-z, arXiv:1609.05327. https://arxiv.org/abs/1609.05327
10. A. Billoire, A. Maiorano, E. Marinari, V. Martin-Mayor, D. Yllanes, 'The cumulative overlap distribution function in realistic spin glasses', Physical Review B 90, 094201 (2014). DOI: 10.1103/PhysRevB.90.094201, arXiv:1406.1639. https://arxiv.org/abs/1406.1639